B5 · Generative AI and Conversational Assistants

Identifying AI-Generated Content

AI-generated content is text, images, audio or video produced by a generative AI tool. It can sometimes be recognised by tell-tale flaws, but these are disappearing, so the more reliable approach is to check where the content came from and whether trustworthy sources confirm it.

8 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

A photo shows a famous cricketer shaking hands with someone he has never met. It looks completely real. Is there any way to tell?

AI-generated content is text, images, audio or video produced by a generative AI tool. It can sometimes be recognised by tell-tale flaws, but these are disappearing, so the more reliable approach is to check where the content came from and whether trustworthy sources confirm it.

In simple words

Forged banknotes used to be easy to spot by their blurry printing. As forgers improved, people stopped relying on a glance and began to check the security thread and the watermark. Generated pictures are following the same path. Looking for mistakes still helps sometimes, but the safer question is no longer "does this look fake?" It is "where did this come from?"

Why it matters

Realistic fake images, voices and videos are used to cheat people of money, to damage reputations and to mislead voters. The opposite danger is growing too: genuine evidence is dismissed as "probably AI". You need a way of judging content that does not depend on your eyes alone.

Key terms

AI-generated content
Text, images, audio or video made by a generative AI tool.
Deepfake
Realistic fake audio or video of a real person, made with AI.
Provenance
The recorded history of a piece of content: who made it, with what, and how it was changed.
Watermark
A mark, sometimes invisible, added to content to show that it was made by AI.
Detector
A tool that estimates whether content was generated by AI. Detectors are often wrong.
Liar's dividend
The advantage a dishonest person gains when real evidence can be dismissed as fake.

Explained in detail

Clues in images

Image generators build a picture from patterns, with no understanding of how objects are put together. This used to leave obvious marks: hands with six fingers, earrings that did not match, spectacles that merged into the face, writing on signboards that dissolved into nonsense, a background where a railing led nowhere. Some of these still appear. Lighting may fall from two directions at once. Skin may look too smooth, like polished wax. A crowd in the background may have faces that melt together. Reflections in a window or a mirror may not match the scene.

Clues in audio and video

In generated video, watch the edges and the movement. Lips may not quite match the words. Blinking may be too rare or too regular. Hair, fingers and teeth may shimmer or change shape between frames. Objects may pass through each other. In generated or cloned audio, listen for a rhythm that is too even, breaths in odd places, a flat emotional tone, or a voice that mispronounces a name the real person would know well. A cloned voice on the telephone often avoids answering unexpected questions.

Clues in text

Generated text tends to be smooth, balanced and general. It may use the same sentence patterns again and again, give neat lists of three, hedge every statement, and contain no personal detail, no specific place or date, and no opinion that anyone could disagree with. These are tendencies, not proof. Many people write in exactly this way, and a person can easily edit generated text to remove such signs.

Why clues are not enough

Every flaw on these lists is something the makers of the tools are working to remove, and each new version removes more. In tests, people asked to tell real faces from generated ones often do little better than tossing a coin, and they tend to be confident when they are wrong. Software detectors do not solve the problem. Detectors for text have wrongly accused students of using AI, with writers whose first language is not English flagged more often. Detectors for images can be defeated by resizing or re-saving the picture. A detector's verdict should never be treated as proof.

The better question: where did it come from?

  • Who posted it first? An established news organisation, or an account created last week?
  • Do reliable outlets report the same event? A real incident involving a famous person is reported by many.
  • Does a reverse image search show the picture existing earlier, perhaps in a different setting?
  • Does the person shown have an official account that confirms or denies it?
  • Does it ask you to do something urgently: pay, click, forward, be outraged?

Labels built into the content

Because detection after the fact is unreliable, attention has turned to marking content at the moment it is made. One approach, known as Content Credentials, attaches a signed record to a file saying which camera or tool made it and how it was edited; the record can be inspected by anyone. Another embeds an invisible watermark into generated images, audio or text that special software can read. Several countries, including India, have introduced or proposed rules requiring AI-generated media to be labelled. These measures help, but they are not complete: not every tool takes part, and labels can be stripped out. The absence of a label proves nothing.

The liar's dividend

There is a second harm that is easy to overlook. Once everyone knows that convincing fakes exist, a person caught on a genuine recording can simply say, "That is AI." Real evidence loses its force. Legal scholars named this the liar's dividend. It means that careless doubt is as damaging as careless belief. The aim is not to distrust everything. It is to check, and then to believe what the evidence supports.

How it works

ContentPossible cluesMore reliable check
ImageOdd hands, garbled writing, waxy skin, mismatched reflectionsReverse image search; who published it first
VideoLips out of step, strange blinking, shimmering edgesReports from reliable outlets; the person's official account
AudioEven rhythm, flat feeling, avoids unexpected questionsCall back on a known number; ask a private question
TextSmooth, general, repetitive, no personal detailCheck the facts in it; ask the author about it

Try it yourself

Is this photo real?

A dramatic photo of a flooded railway station is spreading in a family group.

A clue, or a reliable check?

Clues can mislead. Checks of origin are stronger.

  • A reverse image search shows the photo is five years old

  • The essay sounds very smooth

  • The hand in the picture has six fingers

  • A detector website says "87% AI"

  • Three reliable news outlets report the same event

  • You call your cousin back on her usual number and she says she never rang

Common mistakes

  • Believing you can always tell by looking. The best fakes have no visible flaws.
  • Treating a detector's result as proof, or accusing someone because of it.
  • Dismissing real evidence as "probably AI" without checking.

Check yourself

Question 1

1. Why are visual clues such as odd hands becoming less useful?

Question 2

2. What is the most reliable way to judge a suspicious photo?

Question 3

3. What is the "liar's dividend"?

Summary

  • Generated content sometimes shows flaws, but these are fading and detectors are unreliable.
  • The stronger test is origin: who published it, and do reliable sources confirm it?
  • Avoid both careless belief and careless doubt. Check, then decide.

Finished reading?

Sources: Content Credentials (C2PA) · Civic Online Reasoning (Digital Inquiry Group)